Download SATELLITE MONITORING OF CURRENT AND HISTORICAL DEVELOPMENT by

SATELLITE MONITORING OF CURRENT AND HISTORICAL DEVELOPMENT
PATTERNS IN BIG SKY, MONTANA: 1990-2005
by
Natalie Monique Campos
A thesis submitted in partial fulfillment
of the requirements for the degree
of
Master of Science
in
Land Resources and Environmental Sciences
MONTANA STATE UNIVERSITY
Bozeman, Montana
May, 2008
© COPYRIGHT
by
Natalie Monique Campos
2008
All Rights Reserved
ii
APPROVAL
of a thesis submitted by
Natalie Monique Campos
This thesis has been read by each member of the thesis committee and has been
found to be satisfactory regarding content, English usage, format, citation, bibliographic
style, and consistency, and is ready for submission to the Division of Graduate Education.
Dr. Rick L. Lawrence
Approved for the Department of Land Resources and Environmental Sciences
Dr. Jon M. Wraith
Approved for the Division of Graduate Education
Dr. Carl A. Fox
iii
STATEMENT OF PERMISSION TO USE
In presenting this thesis in partial fulfillment of the requirements for a
master’s degree at Montana State University, I agree that the Library shall make it
available to borrowers under rules of the Library.
If I have indicated my intention to copyright this thesis by including a
copyright notice page, copying is allowable only for scholarly purposes, consistent with
“fair use” as prescribed in the U.S. Copyright Law. Requests for permission for extended
quotation from or reproduction of this thesis in whole or in parts may be granted
only by the copyright holder.
Natalie Monique Campos
May 2008
iv
ACKNOWLEDGEMENTS
I would like to extend my deepest gratitude to my major advisor, Rick Lawrence,
for his constant support and unlimited patience. I would also like to thank Brian
McGlynn and Kathy Hansen for the guidance and input. I am especially thankful to
Brian McGlynn and Kirstin Gardner for providing the funding for this project and for
allowing me to contribute to their research program. I would also like to thank EPA
STAR Understanding Ecological Thresholds in Aquatic Systems through Retrospective
Analysis – Grant #R832449, Seed funding NSF - Geography and Hydrology Programs
(joint funding) - ALSM high resolution topography data acquisition - BCS 0518429,
Montana Department of Environmental Quality - Science to inform the TMDL process,
and USGS 104b Montana seed grant program for their funding. I also want to extend my
gratitude to Anne Loi for saving my data from numerous system failures. Most of all, I
want to thank my three little munchkin angels for being the driving force behind my
ambition.
v
TABLE OF CONTENTS
1. INTRODUCTION ......................................................................................................1
2. LITERATURE REVIEW ...........................................................................................3
Mountain Resort Development ...................................................................................3
High-Resolution Imagery Classification......................................................................5
Object-Oriented Classification ....................................................................................6
Image Fusion ............................................................................................................11
Moderate Resolution Change Detection Methods......................................................13
Summary ..................................................................................................................16
3. OBJECT-ORIENTED LAND COVER/LAND USE MAPPING OF
MOUNTAIN RESORT DEVELOPMENT ...............................................................18
Introduction ..............................................................................................................18
Methods....................................................................................................................20
Results......................................................................................................................26
Discussion ................................................................................................................32
Object Segmentation..............................................................................................32
The Quickbird Classification..................................................................................36
The Fused Classification ........................................................................................38
Conclusion................................................................................................................40
4. LAND USE /LAND COVER CHANGE IN BIG SKY, MT: 1990-2005...................42
Introduction .............................................................................................................42
Methods...................................................................................................................45
Summary of Methods............................................................................................46
Results.....................................................................................................................52
Classification .........................................................................................................52
Change Detection...................................................................................................54
Spatial Pattern Analysis .........................................................................................55
Discussion ...............................................................................................................61
Multi-resolution Image Classification ...................................................................61
Analysis of Temporal Change ...............................................................................63
Conclusion...............................................................................................................65
5. CONCLUSION ........................................................................................................67
LITERATURE CITED..................................................................................................71
vi
LIST OF TABLES
Table
Page
3.1 Classification Hierarchy ....................................................................................24
3.2: Error Matrix for Quickbird Classification Level One.........................................27
3.3: User’s and Producer’s Accuracies for Quickbird Classification Level One ........27
3.4: Error Matrix for Quickbird Classification Level Two ........................................28
3.5: User’s and Producer’s Accuracies for Quickbird Classification Level Two .......29
3.6: Error Matrix for Fused Classification Level One ...............................................29
3.7: User’s and Producer’s Accuracies for Fused Classification Level One ..............30
3.8: Error Matrix for Fused Classification Level Two ..............................................31
3.9: User’s and Producer’s Accuracies for Fused Classification Level Two..............31
4.1: Table of classification scheme for years 1990, 2005, and change image .............51
4.2: Error matrix for 1990 Landsat TM classification ...............................................53
4.3: User’s and producer’s accuracy for 1990 Landsat TM classification..................54
4.4: Percentage of each change class to the total amount of change between 19902005 .................................................................................................................56
4.5: Land cover percentages in 1990 Landsat classification......................................56
4.6: Mean value for indicator variables for from-to change change classes ...............56
4.7: Table of mean and standard deviation values for the indicator variable slope,
for forest and grasslands. ..................................................................................57
4.8: Table of mean and standard deviation values for the indicator variable
distance-to-roads for forest and grasslands .......................................................57
4.9: Table of mean and standard deviation values for the indicator variable
distance-to-stream for forest and grasslands......................................................58
vii
LIST OF TABLES – CONTINUED
Table
Page
4.10: Proportion of forest and grassland for nine different aspects to the overall
amount of forest and grassland in their classification .......................................58
viii
LIST OF FIGURES
Figure
Page
3.1: Study area ............................................................................................................21
3.2: Difference between fused and Quickbird classifications ......................................33
3.3: Comparison of Quickbird false color composite and fused classifications
showing mixed land cover objects.......................................................................35
3.4: Fused classification showing effects of subsetting images due to memory
limitations ...........................................................................................................36
3.5: Quickbird false color composite and Quickbird classification level Two showing
same area ............................................................................................................37
3.6: Comparison of false color composite and fused classification...............................38
4.1: Study area ............................................................................................................46
4.2: Classified 1990 Landsat TM image based on 2005 Quickbird classification .........53
4.3: Classified NDVI difference image showing temporal from-to change classes.......55
4.4: Classification tree results .....................................................................................60
4.5: Classification tree based classified map of from-to changes .................................61
ix
ABSTRACT
The goal of this study was to map current and historical development patterns in
Big Sky, Montana. Object-oriented classifications of a high-resolution Quickbird image
and a fused Quickbird and LiDAR image were compared. Results demonstrated that
object-oriented classification can be used to overcome the difficulty associated with
pixel-based classification of high-resolution images through the addition of contextual
metrics to the classification process. The fused classification resulted in decreased errors
of commission and omission for each class, but the differences between the
classifications were not statistically significant. The fused classification represented the
shapes of land cover objects more precisely based on visual assessment.
Temporal analysis of land cover patterns was accomplished successfully by using
a generalized version of the fused classification to map historical development. Previous
research on multitemporal mapping of multiresolution images has been lacking. Our
research showed that the generalization of a high-resolution classification can be used as
training data for a historical image. Normalized Difference Vegetation Index (NDVI)
image differencing and boosted classification trees were used to identify and classify
areas of change. This resulted in the successful identification of temporal changes in land
cover due to Mountain Resort Development (MRD).
Statistical pattern analysis revealed correlations between MRD and the variables
distance-to-streams, distance-to-roads, slope, and aspect. Forest changes were found to
be disproportionately located farther away from streams and on lower slopes. Grassland
changes disproportionately occurred closer to steams, but overall grassland change was
proportional to grassland land cover in 1990. Classification tree analysis indicated the
variables distance-to-streams, distance-to-roads, slope, and aspect explained 87% of the
variance for the change classes and might be related to amenity development. There was
an increase in impervious surfaces and a decrease in both forests and grassland areas
between the years 1990-2005. Loss of forest and grassland area can result in increased
habitat fragmentation and can have negative consequences for ecosystems within the
areas. Overall, this project successfully mapped both current and historical development
patterns in Big Sky, Montana. This allowed for statistical pattern analysis of variables
that have been shown to be correlated with MRD.
1
CHAPTER 1
INTRODUCTION
Mountain resort development (MRD) is rapidly increasing throughout the
Intermountain West. Population growth in the Greater Yellowstone Ecosystem (GYE),
for example, has increased 55% between 1970 and 1997 (Hansen et al., 2002). Studies
have shown that the main attraction to the intermountain west is the quality of life
associated with living near areas rich with natural amenities (Williams and McMillan,
1983; Williams and Jobes, 1990).
The emerging pattern of development is a shift from a primary extractive
economy to a tertiary economy. Rural land in the intermountain west is now being
converted to residential land for bedroom communities for large mountain resorts and for
“ranchette” type development with mountain scrubland and forested areas being most
often affected by residential development (Riebsame et al., 1996; Odell et al., 2003). The
result of MRD is changes in land use and land cover (LULC), which can affect
ecosystems within the developed area, adjoining undeveloped areas, and downstream and
riparian systems
The location of MRD relative to wild and semi-wild areas can have important
effects on ecological processes and wildlife diversity by causing increased development
of the surrounding region as the population grows and demands for services increase
(Baron et al., 2000). Rapid development often results in environmental degradation
specifically related to waste water disposal (Brown et al., 1997). Despite the possible
2
implications of MRD on water resources, there has been a lack of research in this area
(Shanley and Wemple, 2002).
Remote sensing has the potential to be a valuable tool for analyzing MRD and
changes in LULC over time.
The synoptic nature of remote sensing might enable
efficient mapping of overall MRD patterns. Historical archives of remotely sensed data,
in some cases extending over more than three decades, might enable temporal analysis of
LULC change patterns.
Big Sky, Montana, provides a valuable opportunity to assess LULC associated
with MRD. The first ski resort in the area was established in 1973 with the expansion of
the tourism industry taking place entirely within the history of the United States Land
Remote Sensing Program.
This coincidence in time of development and imagery
archives allows for a full examination of development from inception to the current state
of extensive development and rapid growth.
3
CHAPTER 2
LITERATURE REVIEW
Mountain Resort Development
The possible ecological effects of MRD are numerous and wide ranging.
Wilderness trails and human development have been found to negatively affect bird
communities by changing species composition, resulting in a loss of native species and an
increase in generalized species (Miller et al., 1998; Odell et al., 2003). Native birds have
also been found to nest less often near trails and have a greater predation rate (Miller et
al., 1998).
Tourism pressure increases as mountain resorts become popular destinations,
resulting in greater impacts to the environment. It has been found that recreational use
impacts, such as litter, plant damage, and fire rings, have negative effects on visitor
experiences (Lynn and Brown, 2003). Human activities also change vegetation structure
and distribution which in turn affect animal species richness and abundance (Sauvajot et
al., 1998). The addition of ski runs, for example, results in destroyed vegetation and
increased soil erosion due to vegetation loss (Ries, 1996).
Accessibility has been touted as a primary growth factor for tourism in mountain
areas (Price, 2002). The creation of infrastructure creates accessibility pathways that free
tourism from ecological constraints and allow access into more remote areas furthering
MRD (Price, 2002). The addition of roads and other accessibility pathways increase
4
habitat fragmentation by splitting larger parcels of land into smaller ones, thereby
increasing the overall amount of edge habitats (Reed et al., 1996).
Habitat fragmentation and isolation has been found to be an important predictor of
mammalian predator distribution and abundance (Crooks, 2002). Habitat fragmentation
affects wildlife by decreasing the area needed for home ranges, foraging, and dispersal
patterns (Weaver et al., 1996). Fragmented habitats result in carnivorous mammals
having a higher rate of contact with human settlements than less fragmented areas
(Crooks, 2002).
The development of tourism in mountain areas is not steady overtime. It has been
found that there are varying periods of slow growth often followed by rapid growth
(Price, 2002). On a yearly time scale there is also seasonal variation in tourism that
affects the economy through increased or decreased revenue and employment (Price,
2002). Mountain resorts facilities often need to be maintained year-round, creating jobs
in rural or remote locations and increasing the demand for infrastructure and services.
Development near water has great implications for hydrologic systems. It has
been found that the addition of roads affects watersheds by increasing drainage density
(Wemple et al., 1996). The temporal development of transportation networks has also
been shown to correspond to temporal changes in watershed peak flow patterns (Wemple
et al., 1996). Increases in sediment production due to erosion and deposition processes
have also been attributed to transportation networks (Wemple et al., 2001).
Anthropogenic changes in land cover affect water qualities both in the immediate area
and downstream. Harmful ground water nitrate concentrations have been statistically
5
related to the number of septic tanks found in an area, for example (Gardener and Vogel,
2005).
High-Resolution Imagery Classification
High-resolution imagery provides a detailed look at the land’s surface.
The
Quickbird high-resolution satellite has 2.4-m multispectral resolution at nadir with four
bands located in the visible and near-infrared portion of the electromagnetic spectrum.
This high spatial resolution allows individual objects to be visually recognized, such as a
house or a stand of trees, and is ideal for mapping local areas (Sawaya et al., 2003).
The classification of MRD using high-resolution imagery is difficult due to the
complexity found within the scene. LULC associated with MRD is often characterized
by high density residential and commercial development, fringe suburban development,
outdoor recreation areas such as ski slopes and hiking trails, and natural landscapes such
as alpine forests and rock outcrops.
Urban development is the most complex environment to classify due to its
composition of houses, roads, vegetation, and commercial development, including
structures and parking lots. It has been found that high-resolution sensors lack the
spectral capability to capture within class spectral variance of the urban land cover due to
their large band width and position within the electromagnetic spectrum (Herold et al.,
2003). Investigations on the spectral properties of urban areas found that there were large
amounts of spectral variation between features such as roads and rooftops that can only
6
be distinguished with narrow band intervals (Herold et al., 2004). The spectral variability
between features makes per-pixel classifications based on spectral response difficult.
The high-resolution sensors’ inability to capture the subtle characteristics of the
urban environment results in confusion among classes with similar spectral
characteristics in the visible and near-infrared portions of the spectrum such as water and
asphalt (Sawaya et al., 2003). The spatial detail of high-resolution sensors also adds
shadows from tall objects such as trees and large buildings, which further complicates
classification by obscuring the information within the shadow.
Object-oriented
classification, as opposed to a traditional per-pixel based classification, has been
recommended to overcome classification obstacles (Sawaya et al., 2003, Herold et al.,
2003)
Object-Oriented Classification
Object-oriented classification differs from pixel based classification methods in
that it classifies based on homogenous, spatially contiguous groups of pixels or objects
rather than individual pixels.
The process begins by segmenting an image into
homogenous regions or objects based on user defined parameters such as scale of
heterogeneity, shape, compactness, and smoothness.
The key to the segmentation
process is the creation of semantically meaningful objects, meaning objects resemble
their real world counterparts.
Objects need to “as large as possible and as fine as
necessary” (Definiens Professional 5: LDH, 2007). Once appropriate objects are created
contextual metrics such as shape, texture, and location can be calculated and used in their
7
classification; thus increasing the information extracted from the image (Benz et al.,
2003).
The benefits of using an object-oriented classification are numerous.
The
“simultaneous vector and raster representation” of objects allows for compatibility with
GIS systems (Benz et al., 2003). This allows for the integration of GIS information into
the classification scheme in the form of ancillary data or the exportation of the
classification into GIS systems for further spatial analysis and mapping purposes.
Object-oriented classification has also been found to visually improve classification by
reducing the pixilation effect found with other classification methods (Carleer and Wolff,
2006).
Objected-oriented classification was used to successfully map urban land cover
using high-resolution IKONOS imagery of Santa Barbara, California (Herold et al.,
2002). Land cover classes included: green vegetation, red tile roofs, light tile roofs, dark
roofs, streets/parking lots, swimming pools, natural water bodies, and bare soil/nonphotosynthetic vegetation. Overall accuracy was 79%, with individual class accuracies
of 95%, 96%, and 92% for green vegetation, swimming pools, and water bodies,
respectively (Herold et al., 2002). Low class accuracies for dark roofs and streets/parking
lots, 69% and 68%, were direct results of spectral similarity due to similar composition
(Herold et al., 2002). The ability to obtain any separation between these two spectrally
similar classes was due to the object-oriented classification’s inclusion.
Object-oriented analysis was also used to overcome the spectral similarity of
burned and shadowed areas (Mitri and Gitas, 2004). The object-oriented classification in
8
conjunction with fuzzy logic resulted in overall classification accuracies of 99% for a
topographically-corrected image (Mitri and Gitas, 2004). The high overall accuracy
highlighted the effectiveness of using an object-oriented classification to overcome
spectral similarity.
Object-oriented analysis was used to assess forest conditions of a managed forest
in central Japan (Shiba and Itaya, 1996). This study focused on using high-resolution
multispectral IKONOS data (Bands 2, 3, and 4), a derived NDVI image, and slope and
aspect data derived from a 50-m digital terrain model in the segmentation process (Shiba
and Itaya, 1996). The result of analysis was a segmentation based on forest conditions
for which management practices could be based upon (Shiba and Itaya, 1996).
Forest inventory parameters were extracted using object-oriented image analysis
using IKONOS 2 data in conjunction with Alberta’s Vegetation Inventory data and a 10m x 10-m DEM as ancillary data for a mature forested ecosystem (Chubey et al., 2006).
A decision tree classification algorithm was used to create a land cover classification with
eight classes: pine-dominate, spruce-dominate aspen-dominate shrub land, grassland,
rock, sand, and water (Chubey et al., 2006). The overall accuracy was 93% (Chubey et
al., 2006). Individual class accuracy’s varied from 81% to 100% with spruce dominate
having the lowest accuracy and sand the highest (Chubey et al., 2006).
Object-oriented analysis was used to identify woody vegetation in an urban area
of Phoenix, Nevada using high resolution color photography (Walker and Briggs, 2007).
The object-oriented classification was used to create a binary classification of woody
vegetation (shrubs and trees) and non-woody vegetation (Walker and Briggs, 2007). This
9
study found that in a fine segmentation objects tended to have similar shape prosperities.
This prevented the use of shape characteristics in classification and identification of
woody vegetation (Walker and Briggs, 2007). The texture measures were also similar
across objects and resulted in a classification based solely on the spectral metrics. The
use of object-oriented analysis and classification resulted in an overall accuracy of 81%
(Walker and Briggs, 2007).
Object-oriented and pixel-based classifications were compared in a study which
attempted to map impervious surface areas (Yuan and Bauer, 2006). A Quickbird image
was used to classify five classes: impervious surface, forest, water, non-forested rural,
and shadow, in both the object-oriented and the per-pixel classifications (Yuan and
Bauer, 2006). These classes were then recoded so as to result in a binary classification of
impervious and non-impervious surfaces (Yuan and Bauer, 2006). Results of this study
indicate the object-oriented classification had a 1% increase in accuracy over the pixel
based classification with accuracy of 94% for the binary classification (Yuan and Bauer,
2006). Although the difference between the classification’s accuracy was not substantial,
the five class classification’s overall accuracy was markedly higher with a 93 % over the
pixel-based accuracy of 87% (Yuan and Bauer, 2006).
Object-oriented analysis was also used in a study which attempted to classify
mangrove species composition (Wang et al., 2004).
Three different classification
algorithms were used in the study: a pixel based Maximum Likelihood (ML), a Nearest
Neighbor (NN) found within the Definiens Professional software, and a combined ML
and NN (MLNN) (Wang et al., 2004). The MLNN was performed by merging the
10
spectrally similar classes into one class and performing a ML classification.
The
spectrally similar classes were masked and then imported to the object-oriented software
where it was segmented and classified using the NN algorithm (Wang et al., 2004). The
results of the study indicated that the combined MLNN classification had the highest
accuracy with 91% and a Kappa statistic of 0.81 (Wang et al., 2004). The pixel based
ML had an overall accuracy of 89% with a Kappa of 0.73 (Wang et al., 2004). The NN
had an overall accuracy of 80%, with a Kappa of 0.94 (Wang et al., 2004). The decrease
in accuracy for the NN classification was attributed to the misclassification of one
particular mangrove species, which had a decrease in accuracy of 76% (Wang et al.,
2004). The increase in Kappa is a result in an increase in individual class accuracies for
three mangrove species which had low class accuracies with the ML classification (Wang
et al., 2004).
Classification methods are becoming more complex as the use of object-oriented
analysis becomes more prevalent.
Object-oriented analysis and pixel based
classifications were also compared in a deforestation study of Rondonia, Brazil (Budreski
et al., 2007). The Cart ® 5.0 and a k-Nearest Neighbor (k-NN) algorithm were used on a
temporal series of Landsat images in order to map three different land uses: primary
forests, cleared, and regrowth (Budreski et al., 2007). Results from this study indicated
that there was no statistically significant difference between the accuracies of the pixel or
segmentation based classifications (Budreski et al., 2007).
Object-oriented analysis has also been used to map benthic habitats using aerial
photography for a portion of Texas’s gulf coast region (Green and Lopez, 2007). The
11
study area was first subset into 6 geographic regions prior to analysis and segmented to
create objects related to different benthic habitats (Green and Lopez, 2007).
Classification of the objects was achieved through the use of classification and regression
tree analysis in the See5 software package (Green and Lopez, 2007).
The overall
accuracy for this project was 74% (Green and Lopez, 2007). Individual classes with low
accuracies were used to identify between class confusion and used to gather more field
reference data to be used in the manual editing of the classification (Green and Lopez,
2007). This resulted in an overall accuracy of 90% (Green and Lopez, 2007).
Image Fusion
The purpose of image fusion is to combine information from different sensors in
order to increase the information extracted (Pohl and Van Genderen, 1998). There are
many examples of image fusion. Panchromatic images have been fused with multispectral images to create images with increased spatial resolution.
Different multi-
spectral sensors, such as Landsat and SPOT, have been fused in order to increase
radiometric resolution. Active and passive sensors, such as LiDAR and multi-spectral
sensors, have been fused to create images with both spectral and elevation information.
The utility of data fusion was examined in a study that compared the accuracy of
fused images to single sensor images (Hyde et al., 2006). That study found fused images
had the highest accuracy when estimating maximum and mean canopy heights (Hyde et
al., 2006). The higher R2 for fused data set is related to the increase in information
provided by the combination of spectral and height information. Fused ETM+ data
12
achieved a higher R2 than fused Quickbird data due to ETM+’s additional bands and a
wider range within the electromagnetic spectrum. This allows ETM+ to contribute more
information to the classification.
LiDAR and multi-spectral data fusion was also used to investigate accuracy
differences between fused and individual sensor data sets for estimating canopy height
and stem density (McCombs et al., 2003). Results showed that live individual tree
identification for low-density plots had a higher accuracy for the fused image than
LiDAR alone (McCombs et al., 2003). High-density live individual tree identification
plots also had a higher accuracy for the fused data set than LiDAR alone, but were
substantially less accurate than the low-density plots (McCombs et al., 2003). Tree
height was best represented by LiDAR, which underestimated tree height by 0.15-m on
average for high-density and 0.38-m for low-density, while the fused data underestimated
tree height by 0.42-m for high and 0.5-m for low-density plots (McCombs et al., 2003).
The combination of LiDAR and hyper-spectral data also has been used to create a
thematic map related to the National Vegetation Classification for Woodlands and Scrubs
in the United Kingdom (Hill and Thomson, 2005). The map was created by combining
the first two principal components of 12 selected bands from the HyMap sensor and a
digital canopy height model derived from LiDAR first returns (Hill and Thomson, 2005).
The resulting image was then segmented into parcels and used to create a thematic map
with 10 distinct classes relating to species composition and canopy height that could not
have been produced using any one sensor (Hill and Thomson, 2005). The combination of
13
the LiDAR canopy height model and the two principal components allowed the subtle
differences between vegetation types to be captured and accurately classed.
The utility of data fusion has not been limited to vegetation studies. It has been
used in conjunction with color infrared aerial images to extract buildings and trees from
urban environments (Haala and Brenner, 1999). Data was fused by using a normalized
digital surface model from the laser altimeter data as an additional band in classification
resulting in each pixel having height information. This allowed height attributes to be
used in classification, increasing the accuracy of extracted buildings and trees (Haala and
Brenner, 1999).
Fused data sets have also been used with object-oriented classifications. One
study used a pan-sharpened Quickbird image and object-oriented classification in order to
map fractional vegetation cover of urban environments (Moeller, and Blaschke, 2006).
The original 2.4 multispectral pixels where merged with the 0.6 panchromatic band via
the IHS (Intensity – Hue – Saturation) algorithm (Moeller and Blaschke, 2006). This
resulted in an image with 0.61 spatial resolutions and the spectral information of the 4
original bands (Moeller and Blaschke, 2006). The resulting object-oriented classification
resulted in an overall accuracy of 83% (Moeller and Blaschke, 2006).
Moderate Resolution Change Detection Methods
Change detection is the processes of identifying temporal changes in a pixel’s
digital brightness value and associating the change with corresponding biophysical
14
changes in land cover. Change detection algorithms are as numerous as possible change
detection applications and vary in complexity.
Image differencing is a relatively simple change detection algorithm.
It is
performed by subtracting an image of one date from an image of another date, resulting
in an image with no change centered on the mean and change identified in the tails of the
distribution. Image differencing is one of the most commonly used change detection
techniques and has been recommended for binary change/non change purposes (Lu et al.,
2003). The simplicity of image differencing is its main advantage, and when used with
vegetation indices the spectral response of features is emphasized (Lu et al., 2003). The
disadvantage of image differencing is that it does not provide from-to information and
like CVA requires the selection of a threshold in order to capture change (Lu et al.,
2003). Image differencing was used successfully to identify areas of degradation and
regeneration in the tropical forests of Venezuela using the Modified Soil Adjusted
Vegetation Index (MSAVI) (Guerra et al., 1998). It was found that MSAVI differences
between the two dates were a result of vegetation changes due to swidden agricultural
practices (Guerra et al., 1998).
Principal Component Analysis (PCA) has also been recommended for binary
change/non-change purposes (Lu et al., 2003). A significant advantage of PCA is its
ability to reduce band redundancy (Lu et al., 2003). A significant disadvantage of PCA is
that it is data dependent and can be difficult to interpret and classify (Lu et al., 2003).
Principal Component Analysis (PCA) has been used to monitor land use change due to
rapid urban expansion in China’s Pearl River Delta (Li and Yeh, 1998). Change was
15
identified by performing PCA on a radiometric and geometrically corrected multi-date
stacked image (Li and Yeh, 1998). Final analysis was performed by using a maximum
likelihood classification on the first four principal components, which contained 97% of
the variance (Li and Yeh, 1998).
Change Vector Analysis (CVA) identifies and quantifies change through the
magnitude and movement of a pixel in spectral space (Johnson and Kasischke, 1998).
This results in the creation of unique and discreet change vectors that can then be
categorized as change from-to.
CVA requires preprocessing, such as radiometric
normalization and accurate image-to-image registration in order to avoid spurious change
identification (Johnson and Kasischke, 1998). A significant advantage of CVA is its
ability to process any number of bands. CVA has been successfully used to monitor
forest disturbance and re-growth in the Great Smoky Mountain by monitoring pixel
movement and magnitude of change in tasseled cap feature space (Allen and Kupfer,
2000). CVA has also been used to monitor LULC changes related to vegetation dynamic
in the GYE) over a period of 20 years (Paramenter et al., 2003).
Many studies have compared the accuracy and effectiveness of different change
detection methods. One study compared the ability of four different change detection
methods, image differencing, image regression, tasseled cap differencing and a chi square
transformation, for identifying change from no change (Ridd and Liu, 1998).
The
different methods were applied band by band resulting in 15 change images (Ridd and
Liu, 1998).
The chi square transformation uses information from the six Landsat
reflective bands and resulted in a single change image (Ridd and Liu, 1998). The 16 total
16
change images were compared and results indicated differencing Landsat’s red band had
the highest accuracy (Ridd and Liu, 1998). Results also indicated image differencing and
image regression were both effective and preformed similarly when using the visible
bands to identify change (Ridd and Liu, 1998).
The use of Artificial Neural Networks (ANN) for change detection has also been
investigated (Liu and Lathrop, 2002). ANNs are machine learning algorithms which like
other classifiers, require the use of training and test data in order to extract accurate
information from the input data. This study focused on the comparison of ANNs and
post-classification comparison as well as an investigation on the optimal input for the
classification of from-to urban change (Liu and Lathrop, 2002). Six bands (excluding
thermal) from each of two anniversary date Landsat images were used as inputs for one
classification while the first three principal components from each date were used in a
separate classification (Liu and Lathrop, 2002). Results indicate that classification based
on the principal component input had the highest accuracy (Liu and Lathrop, 2002).
When comparing the use of ANNs and post-classification comparison ANN had a higher
accuracy than post-classification comparison, mainly due to compound error associated
with post classification comparison (Liu and Lathrop, 2002).
Summary
MRD in Big Sky, Montana can be attributed to the creation of mountain resorts
and the attraction of living in an area rich with natural amenities.
LULC changes
associated with MRD affect the native flora and fauna of the area and can affect water
17
quality and other physical processes.
The use of object-oriented classification in
conjunction with high-resolution satellite imagery and LiDAR might be used to create
detailed LULC maps to inventory the current distribution of LULC in Big Sky, Montana.
Change detection methods might then be used to assess change in the historical
distribution of LULC over time.
18
CHAPTER 3
OBJECT-ORIENTED LAND USE/LAND COVER MAPPING OF MOUNTAIN
RESORT DEVELOPMENT
Introduction
Mountain resort development (MRD) is rapidly increasing throughout the
intermountain West (Hansen et al., 2002). MRD generally results in land use and land
cover (LULC) change, which can affect ecosystems within the developed area, adjoining
undeveloped areas, and downstream and riparian systems. The location of MRD relative
to wild and semi-wild areas can have important effects on ecological processes and
wildlife diversity (Baron et al., 2000).
MRD has been found to be spurred by the creation of transportation networks
(Price, 2002).
Transportation networks can negatively affect watershed processes
(Wemple et al., 1996; Wemple et al., 2001). Roads and wilderness trails have been found
to affect species composition and increase habitat fragmentation (Miller et al., 1998;
Reed et al., 1996). Habitat fragmentation is an important predictor of species distribution
and abundance (Crooks, 2002).
MRD needs to be accurately mapped in order to assess and monitor the quantity
and extent of MRD. The use of object-oriented classification and analysis of highresolution imagery is advantageous for mapping land use/land cover (LULC) associated
with MRD. High-resolution imagery provides a detailed view of the land’s surface and is
important for mapping local areas (Sawaya et al., 2003).
Traditional pixel-based
19
classification of high-resolution imagery has been difficult due the lack of spectral depth
of most high-resolution sensors (Herold et al., 2003; Herold et al., 2004). Object-oriented
classification and analysis can overcome pixel-based classification problems by allowing
users to classify based on contextual information extracted from an image in addition to
spectral characteristics (Benz et al., 2003).
Many recent studies have investigated the use of object-oriented analysis for
LULC classification.
The spectral similarity of burned and shadowed areas was
successfully distinguished through the use of object-oriented classification with an
accuracy of 99% (Mitri and Gitas, 2004). The high overall accuracy highlighted the
effectiveness of object-oriented classification in overcoming spectral similarity. Objectoriented classification has also been used to temporally monitor vegetation changes in the
southwestern United States (Laliberte et al., 2004). Results indicate that the addition of
spatial information derived from object-based analysis approximates human logic but
benefits through the addition of automation (Laliberte et al., 2004). Object-oriented
analysis has also been used to identify woody vegetation in urban areas based solely on
mean spectral values (Walker and Briggs, 2007). The spatial or contextual information
was not used in the classification process due to objects having no distinct shape pattern
or texture. The segmentation process in that study, however, did create homogenous
objects and resulted in an overall accuracy of 81%.
Image fusion might also increase the accuracy of object-oriented classification
and analysis. The purpose of image fusion is to combine information from different
sensors in order to increase the information extracted (Pohl and Van Genderen, 1998).
20
The fusion of LiDAR and multispectral images has been shown to increase the accuracy
of forest parameter estimations as compared to single sensors (Hyde et al., 2006;
McCombs et al., 2003). Fused LiDAR and hyper-spectral data have also been used to
map vegetation species composition and distribution (Hill and Thomson, 2005). The
utility of data fusion has not been limited to vegetation studies.
The extraction of
buildings and trees from urban environments has been accomplished through the use of
fused aerial photographs and laser altimeter data, which allowed height to be used in
classification and increase the accuracy (Haala and Brenner, 1999).
The purpose of this study was to evaluate the utility of high-spatial-resolution
imagery fused with airborne laser swath mapping, herein refered to as LiDAR, to map
LULC associated with MRD using object-oriented analysis. Results were compared with
object-oriented analysis in the absence of LiDAR data.
Methods
The study area for this project was the West Fork of the Gallatin River watershed
near Big Sky, Montana (Figure 3.1). The study area comprised 24091.3 ha. Big Sky is
surrounded by the Gallatin National Forest in southwestern Montana and is located
within the Greater Yellowstone Ecosystem (Marston and Anderson, 1991). Elevation
ranges more than 2000-m and is an important predictor of climate and vegetation species
distribution (Marston and Anderson, 1991).
Vegetation is composed of coniferous
forests, shrublands, and grasslands. Frost free days range from 60-90 and decrease with
increased altitude (Parmenter et al., 2003; Marston and Anderson, 1991).
21
Figure 3.1: Location of West Fork of Gallatin River watershed study area to the state of
Montana. The study area boundary is shown in black and the West Fork tributaries are
shown in blue.
Imagery used for the study included an 11-tile Quickbird image with 2.4-m spatial
resolution. A tile represents the 16.5 km x 16.5 km footprint of the sensor containing
only the portion of the scene related to the study area with all outlying areas containing
no data. The Quickbird image contained 4 bands in the visible and near infrared (NIR)
portion of the electromagnetic spectrum including blue (450-520 nm), green (520-600
nm), red (630-690 nm), and NIR (760-900 nm). The 11 tiles were mosiacked to create
one master image.
22
1-m LiDAR bare earth and surface models created from point clouds were also
used. The bare earth model was created through triangulation of the last returns in a 55m window (NCLAM processing report, 2005). The surface model was created using a
linear Kriging algorithm to interpolate first returns in a 5-m window. Both models were
exported to ESRI Arcinfo GRID format. The surface model represented surface land
cover, such as the top of tree canopies and the roofs of buildings. The bare earth model
estimated elevation with all land cover (e.g., vegetation and buildings) removed. The
Lidar images were resampled to 2 m using nearest neighbor resampling. All images,
Quickbird and LiDAR, were acquired in the summer of 2005 and registered to a UTM
NAD 83 Zone 12 projection with an RMSE of 0.06.
The original 11 tiles of Quickbird imagery were used to create six image subsets.
Subsetting was necessary due to data processing limitations within the object-oriented
software. The six subsets were then used to create matching subsets of the resampled
LiDAR bare earth and surface models.
The subsets were then imported into the
Definiens Professional 5: Large Data Handling (LDH) software (Definiens Professional
5: LDH, 2007).
The Quickbird and LiDAR fused data created data sets with
multispectral, bare-earth elevation, and surface information for each pixel.
The subsets were initially segmented using a Multi-Resolution Segmentation
(MRS) algorithm (Definiens Professional 5: LDH, 2007). Separate segmentations were
created for each subset with and without LiDAR data.
The MRS algorithm is a
heuristically applied algorithm that creates objects by minimizing internal object
heterogeneity. Object heterogeneity is calculated as a weighted average across all input
23
bands with the total weight summing to one. Object heterogeneity is controlled by
selecting an arbitrary scale factor, which determines the amount of heterogeneity
acceptable and is resolution dependant. The scale factor required to create objects for
individual houses will be different than the scale factor required to create objects
representing a neighborhood or subdivision, for example. The appropriate object is one
that is “as large as possible and as fine as necessary” and thus will vary depending on the
desired output (Definiens Professional 5: LDH, 2007).
The segmentation process
resulted in vector based “objects” with attributes corresponding to the mean and standard
deviation values of the pixels within the object for each input layer.
Additional
contextual metrics can then calculated and used in the classification process.
The
appropriate MRS was determined if objects borders did not overlap different LULC
classes by visual inspection.
The creation of objects is also influenced by weighting layer pixel values and
spatial homogeneity. The default was a pixel value weight of 0.9 and a spatial weight of
0.1 (Definiens Professional 5: LDH, 2007). A decrease in pixel input value weight and a
corresponding increase in spatial weight results in objects with less similarity in their
pixel values. The default spectral and spatial weights were chosen because they provided
objects that followed spectral contrast lines for different land cover by visual inspection.
A Spectral Difference Segmentation (SDS) was performed after the appropriate
MRS was identified (Definiens Professional 5: LDH, 2007). SDS is a merging algorithm
designed to merge spectrally similar objects produced in previous segmentations.
Objects were merged if their standard deviation was below a user defined threshold. The
24
threshold is used to determine the amount of object aggregation. The appropriate SDS
was identified if most or all adjacent objects were merged without creating mixed land
cover objects. This resulted in larger objects, which were more semantically meaningful.
A hierarchical classification scheme was used in this study (Table 3.1). This
study was part of a larger study examining the effect of MRD on stream nitrogen levels,
and the primary interest was to distinguish impervious surfaces from other specified land
cover types. We sought to determine in addition, however, whether our approach could
also distinguish various impervious surface types. Waste water holding ponds and golf
courses represent thematic areas that were manually digitized at the end of the
classification process.
Table 3.1: Classification hierarchy for fused and non-fused images level one and level
two.
Level One
Level Two
Roads
Impervious Surfaces
Bare Soil
Golf Course
Vegetation Not Tree
Lake/Pond
Waste Water Holding Pond
Shadow
Snow
Tree
River/Stream
Building
Rock
Bare Soil
Golf Course
Vegetation Not Tree
Lake/Pond
Waste Water Holding Pond
Shadow
Snow
Tree
River/Stream
Classification of objects was conducted using the nearest neighbor (NN)
algorithm in Definiens Professional. The NN algorithm classifies objects based on user
25
identified sample objects utilizing user selected metrics. Metric selection was determined
using Definiens’ Feature Space Optimization tool (FSO). FSO works by examining any
number of input variables and identifying the variables that contain the greatest distance
between samples to be applied to the NN classifier.
The following pixel derived metrics were selected by FSO: object mean and
standard deviation of all input bands excluding blue (green, red, NIR, NDVI, surface, and
bare earth). NDVI is a commonly used vegetation index that is calculated as (NIRred)/(NIR+red) (Rouse et al., 1973).
NDVI is a unitless measure with a positive
correlation to vegetation amount or health. The following contextual metrics were used:
length/width, asymmetry, density, compactness, and rectangular fit.
Training samples were chosen by selecting objects in each image that represented
their class. Additional training objects were added to improve misclassified classes and
new classifications were conducted until improvements in classification rates were
minimal. Definiens Professional allowed the sample metrics collected in one image to be
applied to subsequent images. This allowed the sample metrics collected in the first
subset to be applied in the classification of subsequent subsets. Once all 12 subsets were
classified, they were mosaicked to create two final classifications, a Quickbird
classification and a fused LiDAR and Quickbird classification.
The fused classification required additional post processing due to missing data
values within the LiDAR image.
classified values to the missing data.
This was overcome by applying the Quickbird
26
The vector based classification was converted to a raster image with 2.4-m pixel
resolution to retain the spatial properties of the original Quickbird data. A total of 1126
accuracy assessment points were generated using a stratified random sample from the
Quickbird classification.
An additional 609 stratified random sample points were
obtained from the fused classification. The two sets of points were merged and used to
assess the accuracy of both classifications. Producer’s and user’s accuracies for each
class and the Kappa statistics were calculated for each classification (Congalton and
Green, 1999). Overall accuracy was calculated using the methods outlined in Carrao et
al., 2007.
This method differs from the traditional overall accuracy, in that it is
calculated based on the relative proportion of each class to the total number of classified
pixels.
Results
The overall accuracy for the Quickbird classification level one was 90% with a
Kappa statistic of 0.78 (Table 3.2).
The bare soil class had the lowest user’s and
producer’s accuracy and was often confused with impervious surfaces (Table 3.3). The
error of omission for bare soil was 72%, meaning that 72% of the bare soil pixels were
not classified as such.
The error of commission relates to what percentage of the
classified pixels was actually the class they were classified. Grass had a high error of
commission rate with 35%, meaning 35% of the pixels were classified as grass but were
not. Grass was confused with impervious surfaces and trees. The river, shadow, and
snow classes each preformed similarly with a low producer’s accuracy and high user’s
27
accuracy. Impervious surface, lakes/ponds, and trees all achieved user’s and producer’s
accuracies over 85%.
Bare Soil
Impervious
Grass
Lake
River
Shadow
Snow
Tree
Row Total
Table 3.2: Error Matrix for Quickbird Classification Level One.
Bare Soil
35
12
2
0
0
0
0
4
53
Impervious
39
589
8
0
0
12
30
8
686
Grass
45
31
168
1
1
0
0
12
258
Lake
0
2
0
46
1
0
0
0
49
River
0
0
0
0
17
0
0
0
17
Shadow
0
0
0
0
0
52
0
0
52
Snow
2
3
0
1
0
0
73
2
81
Tree
6
31
11
4
6
5
2
474
539
Column Total
127
668
189
Overall Classification Accuracy = 90%
52
25
69
105 500 1735
Overall Kappa Statistic = 0.78
Table 3.3: User’s and Producer’s Accuracies for Quickbird Classification Level One.
Reference
Classified
Number
Producer's
User's
Class Name
Pixel Totals Pixel Totals
Correct
Accuracy
Accuracy
Bare Soil
127
53
35
28%
66%
Impervious
668
685
588
88%
86%
Grass
0
1
0
89%
65%
Lake
189
258
168
88%
94%
River
52
49
46
68%
100%
Shadow
69
52
52
75%
100%
Snow
105
81
73
70%
90%
Tree
500
539
474
95%
88%
28
The Quickbird classification level two achieved an overall accuracy of 90% with
a Kappa statistic of 0.76 (Table 3.4).
All individual class user’s and producer’s
accuracies remain the same except for the expansion of the impervious class into
buildings, rocks, and roads (Table 3.4).
The building class had a low error of
commission as seen in the high user’s accuracy but also had a higher error of omission as
seen in the low producer’s accuracy. The road class had a higher error of commission
than omission. The rock class had a high error of commission with 35% and a moderate
error of omission with 11%.
Bare Soil
Building
Grass
Lake
River
Road
Rock
Shadow
Snow
Tree
Row Total
Table 3.4: Error Matrix for Quickbird Classification Level Two.
Bare Soil
35
1
2
0
0
10
1
0
0
4
53
Buildings
0
131
0
0
0
4
1
0
0
0
136
Grass
45
11
168
1
Lake
0
0
0
47
River
0
0
0
0
Road
20
48
0
0
Rock
19
5
8
0
Shadow
0
0
0
0
Snow
2
1
0
1
Tree
6
19
11
4
Column
127
216
189 53
Total
Overall Classification Accuracy = 90%
1
1
17
0
0
0
0
6
12
0
0
202
13
0
0
7
9
1
0
3
181
0
2
5
0
0
0
0
12
52
0
5
0
0
0
0
30
0
73
2
12
0
0
1
7
0
2
474
259
49
17
274
275
52
81
539
25
248
203
69
105
500
1735
Overall Kappa Statistic = 0.76
29
Table 3.5: User’s and Producer’s Accuracies for Quickbird Classification Level Two.
Reference
Classified
Number
Producer's
User's
Class Name
Total
Total
Correct
Accuracy
Accuracy
Building
216
136
131
61%
96%
Road
248
274
202
81%
74%
Rock
203
275
181
89%
66%
The fused classification level one had an overall accuracy of 91% and a Kappa
statistic of 0.8 (Table 3.6).
Bare soil had similar error rates to the Quickbird
classification (Table 3.7). The error of omission for bare soil, however, increased over
the Quickbird classification level one.
All other individual user’s and producer’s
accuracy were generally higher than the Quickbird classification resulting in reduced
errors of omission and commission.
Bare Soil
Impervious
Grass
Lake
River
Shadow
Snow
Tree
Row Total
Table 3.6: Error Matrix for Fused Classification Level One.
Bare Soil
19
2
0
0
0
0
0
0
21
Impervious
57
608
5
0
0
2
14
10
696
Grass
45
37
172
2
0
0
4
19
279
Lake
0
2
0
47
1
0
0
0
50
River
0
0
0
0
21
0
0
0
21
Shadow
0
0
0
2
0
62
0
0
64
Snow
Tree
2
4
6
13
1
11
0
1
0
3
0
5
86
1
2
469
97
507
Column Total
127
668
189
Overall Classification Accuracy = 91%
52
25
69
105
500 1735
Overall Kappa Statistic = 0.8
30
Table 3.7: User’s and Producer’s Accuracies for Fused Classification Level One.
Reference
Classified
Number
Producer's
User's
Class Name
Totals
Totals
Correct
Accuracy
Accuracy
Bare Soil
127
21
19
15%
90%
Impervious
668
696
608
91%
87%
Grass
189
279
172
91%
62%
Lake
52
50
47
90%
94%
River
25
21
21
84%
100%
Shadow
69
64
62
90%
97%
Snow
105
97
86
82%
89%
Tree
500
507
469
94%
93%
Overall accuracy for the fused classification level two was 91% with a kappa
statistic of 0.78 (Table 3.8). The expansion of the impervious class resulted in buildings
having a 16% decreases in error of omission than the Quickbird classification and a slight
1% decrease in error of commission (Table 3.9). The road class had a 5% decrease in
error of commission with the fused classification and no increase in the error of omission.
The rock class had slight decreases in both errors of omission and commission than in the
Quickbird classification.
31
Bare Soil
Building
Grass
Lake
River
Road
Rock
Shadow
Snow
Tree
Row Total
Table 3.8: Error Matrix for Fused Classification Level Two.
Bare Soil
19
1
0
0
0
1
0
0
0
0
21
Building
1
166
0
2
0
4
0
0
0
0
173
Grass
45
12
172
10
0
18
7
0
4
19
287
Lake
0
0
0
37
1
0
1
0
0
0
39
River
0
0
0
0
19
0
0
0
0
0
19
Road
24
27
0
1
0
202
1
0
1
0
256
Rock
32
4
5
0
0
20
184
2
13
10
270
Shadow
0
0
0
2
1
0
0
62
0
0
65
Snow
2
1
1
0
0
0
5
0
86
2
97
Tree
4
5
11
1
Column
127
216
189 53
Total
Overall Classification Accuracy = 91%
4
3
5
5
1
469
508
25
248
203
69
105
500
1735
Overall Kappa Statistic = 0.78
Table 3.9: User’s and Producer’s Accuracies for Fused Classification Level Two.
Reference
Classified
Number
Producer's
User's
Class Name
Totals
Totals
Correct
Accuracy
Accuracy
Building
216
173
166
77%
96%
Road
248
256
202
81%
79%
Rock
203
270
184
91%
68%
Each classification was compared to its counterpart, the fused classification
versus the Quickbird classification, in order to determine if any statistical difference
between the two existed. It was found that neither the level one nor the level two
classifications had any statistical differences in Kappa statistics (Congalton and Green,
1999). The z statistic for the level one classification was 1.13 and 1.06 for level two (pvalues = 0.02), both are below the critical value of 1.96 at an alpha of 0.05 (Congalton
and Green, 1999).
32
Discussion
Object Segmentation
The segmentation process was successful in the creation of semantically
meaningful objects. Houses were easily identified in both segmentations and generally
resulted in independent objects. This allowed contextual information in addition to
spectral and elevation information to be used in the classification. The key for the
segmentation process was the creation of homogeneous objects. This was achieved
through the heuristic nature and application of the MRS and SDS segmentation
algorithms. The results were objects that represented their land cover class spectrally and
contextually.
The use of the surface and elevation models resulted in objects that appeared to
better reflect surface features compared to using Quickbird imagery alone (Figure 3.2).
The addition of elevation information in the fused classification resulted in many more
objects than the Quickbird classification. This allowed smaller patches of land cover to
be identified and objects to have a much more distinguishable shape. Dense housing, for
example, in the Quickbird classification was block shaped, while the same area in the
fused classification had individual houses that were shaped similar to their physical
structure (Figure 3.2).
33
Figure 3.2: Difference between fused (left) and Quickbird (right) classifications.
The issue of scale presented a problem for the segmentation process. Previous
studies have focused on the use of homogenous landscape regions such as primarily
urban or natural areas (Chubey et al., 2006; Laliberte et. al., 2004; Walker and Briggs,
2007; Budreski et al., 2007). These studies have shown the successful use of objectoriented classification and analysis, but have not explored the use for heterogeneous
landscapes. There were two major land cover types within our study area, developed and
undeveloped.
Developed land cover consisted of road networks and buildings.
Undeveloped land cover consisted of grassland, river/stream, lake/pond, and forested
areas. These types have different levels of appropriate segmentation. The appropriate
level of segmentation to create an object representing a house will be smaller than the
level needed for forests or grasslands. An intermediate segmentation scale was used as a
34
compromise in order to create objects “as large as possible and as fine as necessary”
(Definiens Reference Book, 2005, p. 11).
Mixed land cover objects caused difficulty in the selection of class samples.
Samples that represented their class while still capturing the variability of their class were
selected. This required the selection of some mixed land cover objects as samples of the
dominant land cover they contained. The spectral attributes of an object were the means
of all pixels it was composed of for each input. Mixed land cover objects, therefore,
skew the distribution of the response for the dominate land cover they represent. This is
a likely explanation of misclassification rates (Figure 3.3).
Objects that were
predominately impervious also tended to have small patches of vegetation, for example.
Point-based accuracy assessment does not account for mixed land cover objects.
Thus, true accuracy is unknown. Point-based accuracy only accounts for what is on the
ground at one particular point. Points located on the edges of objects or part of a mixed
land cover objects could inflate error rates and lead lower accuracies. New methods of
calculating overall accuracy are needed for full evaluation of object-oriented analysis.
35
Figure 3.3: Comparison of Quickbird false color composite and fused classifications
showing mixed land cover objects.
Transition zones created difficulties in both classifications. Bare earth and grass
are not discrete land cover types, for example. They often flow between each other with
varying intensity. This caused difficulty with the creation of distinct boundaries and
resulted in the low accuracy for the bare soil class.
The object-oriented software was memory intensive and had data processing
limitations. High-spatial-resolution images contain large numbers of pixels for spatially
small areas. The Quickbird full image was 64 MB and the LiDAR image was 1.7 GB.
The software used could not handle the file sizes and therefore required subsetting the
data and resampling the LiDAR data. This created objects that did not always flow
between subsets (Figure 3.4). A study that mapped benthic habitat on the Gulf Coast of
Texas also needed to resample their 1-m aerial photographs and subset their study into
several smaller potions (Green and Lopez, 2007). Other studies have focused on the use
36
of small, less computationally demanding areas (Chubey et al., 2006; Laliberte et al.,
2004; Walker and Briggs, 2007; Budreski et al., 2007; Benz et al., 2003).
Figure 3.4: Fused classification showing effects of subsetting the images due to memory
limitations.
The Quickbird Classification
Overall accuracy for the Quickbird classification was high at 90% for both level
one and level two. Previous studies have found the lack of spectral depth of highresolution sensor makes pixel-based classification difficult (Sawaya et al., 2003). Classes
such as roads, shadows, and water bodies have had high misclassification rates with
traditional pixel-based classification (Sawaya et al., 2003; Herold et al., 2003). Objectoriented classification in our study was able to overcome this through the addition of the
contextual metrics to the classification process.
37
The Quickbird classification had difficulties with several classes.
The bare
ground class, for example, was often misclassified as impervious surface or grass. This
was a result of the continuous nature of bare soil and the creation of discrete boundaries.
Small patches of bare soil were often contained in larger objects of impervious surface or
grass (Figure 3.5). The impervious surface classes were also confused with forest and
grass. Similar to the bare ground class, misclassification can be attributed to small
patches not being identified.
Figure 3.5: Quickbird false color composite and Quickbird classification level two
showing same area.
Previous studies on the use of object-oriented classification have found similar
results. One study mapped densely populated areas of Santa Barbara, California with an
overall accuracy of 79% (Herold et al., 2002). Another study mapped an area of mixed
residential and agriculture land cover with an overall accuracy of 74% (van der Sande et
38
al., 2003). Both studies found that roads or building classes had confusion (Herold et al.,
2002; van der Sande et al., 2003). This was mostly a result of the two classes not being
separated in the segmentation process, as seen in our classification. Our classification
had confusion with bare soil and grass. This was overcome in one study by the use of
broad classes such as non-photosynthetic vegetation/bare soil and general photosynthetic
vegetation (Herold et al., 2002).
The Fused Classification
The accuracy of the fused classification was slightly higher than the Quickbird
classification, but no significant differences were found. The additional information of
the elevation and surface models resulted in the segmentation creating objects that better
reflected surface objects by adding contrast in areas of spectrally low contrast (Figure
3.6). This can be seen in the creation of objects representing small groups of trees and
small patches of bare soil. The Kappa statistic did not have a statistically significant
increase, but the estimated overall accuracy and Kappa statistic were higher. This is
consistent with our observations that the differences were represented by more accurate
delineation of objects, although the total area affected by these differences was small.
Individual class user’s and producer’s accuracy also improved for many of the classes,
thus reducing the error of commission and omission.
39
Figure 3.6: Comparison of false color composite and fused classification.
The fused classification had similar accuracy difficulties as the Quickbird. Bare
soil had a high rate of misclassification with impervious surfaces and grass. Impervious
surfaces had a high confusion with grass. Snow also had a high misclassification rate
with impervious surfaces. These errors can be attributed to the creation of discrete
boundaries for continuous land cover types. Overall, the fused classification had higher
consistency, as there were less dramatic differences between the user’s and producer’s
accuracies than seen in the Quickbird classification.
Previous studies have used image fusion with topographic data with great success.
Studies have compared image fusion to single sensors and found that image fusion results
in increased accuracy for pixel based classifications due to the inclusion of elevation
information to the classification process (Hyde et al., 2006; McCombs et al., 2003; Hill
and Thomson, 2005). Our results show little improvement in overall accuracy when
40
compared with the Quickbird classification. This is a result of the success and power of
the segmentation processing in creating objects, which can be accurately classified
through the addition of contextual metrics. The addition of the LiDAR resulted in finer
objects, which reduced error rates but did not improve overall accuracy.
Conclusion
Object-oriented analysis and classification works well with high-resolution
imagery. Object-oriented classification can overcome the traditional spectral limitations
of high-resolution sensors. The appropriate segmentation level coupled with the addition
of contextual metrics can accurately create detailed maps. Object-oriented classification
can be used to map accurately heterogeneous land cover areas such as dense urban areas,
rural or suburban areas, and natural land cover.
Image fusion and object-oriented classification creates realistic objects.
This
results in a classification that is more visually appealing. This is advantageous if spatial
precision required, but might not be necessary if only point-based accuracy is needed.
There are a plethora of future research opportunities for object-oriented
classification and analysis. One significant research area includes identification of an
appropriate spatial resolution for use with object-oriented classification and analysis.
Quickbird imagery has high spatial resolution but is still pixilated, which blurs the
boundaries of different land cover. Increased class accuracy and homogenous objects
might be achieved in future research if the optimum spatial resolution was identified.
41
Another research area would be the full utilization of contextual metrics. The
Definiens Professional software allows for an extreme number of metrics to be
calculated, including standard deviations and ratios of both spectral and contextual
metrics. This is a new frontier in image classification, as most of these have not be tested
for relevance or relation to different LULC classes. These new metrics need to be
researched and identified, as has been done for NDVI, for use in the classification of
objects.
Object-oriented software also needs to be programmed so as to make full use of
the today’s high resolution data. Today’s technology is resulting in an increased number
of sensors with increasing spatial resolution. This software’s inability to work with large
data sets relegates it to a novelty procedure not currently useful for landscape-scale
classification.
42
CHAPTER 4
LAND USE /LAND COVER CHANGE IN BIG SKY, MT: 1990-2005
Introduction
Mountain Resort Development (MRD) is rapidly increasing throughout the
intermountain west. The result is changes in land use and land cover (LULC), which can
affect ecosystems within the developed area, adjoining undeveloped areas, and
downstream and associated riparian systems.
The conversion of naturally vegetated land to impervious surfaces, such as roads,
can have dramatic affects. Roads have been found to increase habitat fragmentation
(Reed et al., 1996). Habitat fragmentation in turn has been found to decrease species
composition distribution and abundance (Odell et al., 2003; Crooks, 2002; Miller et al.,
1998).
Roads also affect physical processes.
Roads have been shown to increase
drainage density and sediment production due to erosion and deposition (Wemple et al.,
2001; Wemple et al., 1996).
Understanding the variables associated with potential causes of LULC change due
to MRD is important in order to mitigate its effects. Accessibility has been named as a
primary growth factor for tourism in mountain areas (Price, 2002). We would expect,
therefore, that MRD would be correlated to roads within a watershed.
Other studies have shown that the quality of life associated with living near areas
rich with natural amenities is a significant attraction for development in rural areas
(Williams and McMillan, 1983; Williams and Jobes, 1990). Amenity development often
43
results in the conversion of rural land to residential land for “ranchette” type development
(Riebsame et al., 1996; Odell et al., 2003). Topography and water area are important
variables and have been considered the basis of a “natural amenity index” (Cromartie and
Wardell, 1999).
Evaluation of MRD with respect to topographic variables is also
important because of its effects on watershed processes. Increased development on
steeper slopes, for example, might increase rates of erosion and related impacts of aquatic
ecosystems (Wemple et al., 2001; Groffman et al., 2003).
Remote sensing has the potential to be a valuable tool for analyzing MRD and
changes in LULC over time. The use of remote sensing for identification of LULC
change has been well documented (Coppin et al., 2004; Mas, 1999). Change detection
algorithms are as numerous as possible applications and vary in complexity.
Many studies have compared the accuracy and effectiveness of various change
detection methods. Univariate image differencing has been recommended for binary
change/non-change identification (Coppin et al., 2004; Lu et al., 2003). The use of
vegetation indexes for change detection also has been shown to be advantageous over
single band analysis, because it reduces data volume and captures information not
available in any single band (Coppin et al., 2004). Normalized Difference Vegetation
Index (NDVI) image differencing has been shown to be advantageous over the use of
other vegetation indexes for identification of deforestation and vegetation loss and has
also been shown to be less affected than other indexes by topographic relief (Lyon et al.,
1998). This makes NDVI image differencing an ideal candidate for binary change/no-
44
change identification related to MRD where we are examining the clearing of natural
habitats for anthropogenic development in an area of high relief.
Classification of from-to change is imperative, once change has been identified, in
order to discern development patterns. Classification tree (CT) methods are becoming
increasingly popular for classification of remotely sensed data and might be useful for
such from-to change. CTs have been shown to create more accurate classifications than
other methods (De’ath and Fabricius, 2000). CTs offer the advantages of being nonparametric, working equally well with continuous and nominal data types, producing
interpretable rule sets, and handling noisy data sets (Friedl and Brodley, 1997). CT,
therefore, offers significant advantages over other classification methods; however, they
can be negatively affected by the presence of outliers in training data and by unbalanced
data sets (Lawrence et al., 2004). Statistical boosting has been shown to significantly
increase the accuracy of CTs (Bricklemeyer et al., 2007; Baker 2007; Baker et al., 2006;
Lawrence et al., 2004).
Multi-resolution image classification has been shown to increase single image
classification accuracy (Moeller and Blaschke, 2006; Hyde et al., 2006; McCombs et al.,
2003). Multi-resolution image classification is generally considered a form of image
fusion whose purpose is to combine information from different sensors in order to
increase the information extracted (Pohl and Van Genderen, 1998). The use of multiresolution images for change detection purposes, however, does not seem to be as well
developed as use for single classifications. This raises a significant issue with modern
high-resolution satellite imagery. High-resolution imagery is increasingly being used for
45
LULC classifications, but the absence of a long-term archive of high-resolution imagery,
as exits for moderate-resolution imagery, has made it less useful for LULC change
analysis.
Development of approaches that fuse high-resolution classifications with
moderate-resolution imagery for LULC change analysis would greatly increase the
usefulness of high-resolution classifications.
The purpose of this study was to (1) identify LULC change related to MRD in Big
Sky, Montana, between the dates of July 2005 and July 1990 using high-resolution
imagery from 2005 and moderate resolution imagery from 2005 and 1990 and (2)
evaluate topographical variables and spatial relationships to roads and streams as possible
correlates of MRD in the area.
Methods
The study area for this project was the West Fork of the Gallatin River watershed
near Big Sky, Montana (Figure 4.1). The study area comprised 24091.3 ha. Big Sky is
surrounded by the Gallatin National Forest in southwestern Montana and is located
within the GYE (Marston and Anderson, 1991). Elevation ranges more than 2000 m and
is an important predictor of climate and vegetation species distribution (Marston and
Anderson, 1991).
grasslands.
Vegetation is composed of coniferous forests, shrublands, and
Frost free days range from 60-90 and decrease with increased altitude
(Parmenter et al., 2003; Marston and Anderson, 1991).
46
Figure 4.l: Location of West Fork of Gallatin River watershed study area to state of
Montana. The study area boundary is shown in black and the West Fork tributaries are
shown in blue.
Summary of Methods
Identification of change patterns for statistical analysis was a multistep process.
Two normalized near anniversary date Landsat 5 images were converted to NDVI and
differenced to identify change from no-change. Classification of from-to change was
carried out by first resampling a 2005 Quickbird classification with an accuracy of 91%
to 30 m. Identified change locations in the 1990 Landsat image were classified, using the
unchanged locations as training data.
Changed areas were then combined with the
unchanged locations in the 2005 classification to create the final 1990 classification. The
47
spatial distribution of change was analyzed through a combination of descriptive statistics
and CT.
Landsat 5 Thematic Mapper (TM) scenes (path 39 and row 28) from July 12,
2005 and July 3, 1990 images were obtained from EROS Data Center with level one
terrain correction.
The two scenes were geometrically registered to a July 2005
Quickbird image and set to a UTM NAD 83 Zone 12 projection. The Landsat 1990
image had an RMSE of 0.05 and the 2005 image had a RMSE 0.07. The full scenes were
then subset to the study area.
Data transformations were performed on both Landsat dates to provide derived
indexes for use in the classification and change detection processes. A tasseled cap
transformation was performed. The tasseled cap is a linear transformation of the original
six Landsat reflective bands resulting in three new bands representing relative soil
brightness, relative amount of green vegetation, and relative soil moisture content (Crist
and Cicone, 1984). NDVI was also calculated as (NIR-red)/(NIR+red) (Rouse et al.,
1973). NDVI is a unitless measure with a positive correlation to vegetation amount or
health.
A multi-date image normalization using linear regression was used as a
radiometric normalization procedure on the 1990 and 2005 NDVI images (Collins and
Woodcock, 1996). The 2005 image was selected as the independent variable and the
1995 image was selected as the response variable. Pseudo-invariant features (PIFs)
representing features that were presumed to have not changed over time were identified
and selected. Stable anthropogenic features included features that were both bright (e.g.,
48
buildings) and dark (e.g., man-made water features) and were ideal candidates; an
attempt was made to select features in relatively flat areas so as not to introduce
topographic effects. A total of 224 points were collected. 100 points were randomly
selected and set aside to be used as a validation data set. The R2 for the NDVI regression
model was 0.91 and was shown to be statistically significant (p-values of <0.001).
Paired t-testing was performed in order to determine whether the normalization
improved the 1990 image. It was determined that the mean difference between the 2005
and the 1990 image was 7.4 with a standard deviation of 17.2. The mean difference
between the 2005 and the normalized 1990 image was 6.1 with a standard deviation of
9.3. The paired t-test indicated that there was a difference between both the 1990 image
and the 1990 normalized image and the 2005 image. This difference was decreased,
however, with the normalization process.
NDVI image differencing was used for binary change/no-change identification.
Image differencing was preformed by subtracting the pixel values of the 2005 NDVI
image from the pixel values of normalized NDVI 1990 image. The resulting values have
no-change centered on the mode and areas of change located in the tails of the
distribution. Change was then identified by defining a threshold based on knowledge of
certain changed locations to separate areas of change from no-change. The appropriate
threshold was identified so as to include all areas of change while minimizing false
positive change identification. A binary image of change and no-change was created
based on this threshold.
49
A 30-m resolution 2005 classification was created based on a classified fused
Quickbird and LiDAR classification with a resolution of 2.4 m. The original classes of
grasslands/shrublands,
bare
soil,
and
golf
courses
were
generalized
to
grasslands/shrublands. Houses, roads, and rock were generalized to impervious surfaces.
Lakes, rivers, ponds, streams, and waste water holding ponds were generalized to water.
The forest class was generalized to forest. Snow and shadows were left to represent
themselves.
The generalized 2.4-m classification was then resampled to 30 m to match the
1990 Landsat image. Resampling was accomplished by determining within each 30-m
pixel the most common land cover type present.
The use of the resampled generalized classification resulted in 30-m pixels that
represented mixed land cover. This resulted in the loss of important information related
to MRD, such as roads and other smaller features. A percent impervious model was
created in order to overcome this loss of information by determining within each 30-m
pixel the percent of 2.4-m pixels that were impervious, based on the 2.4-m classification.
Pixels that were 20% or more impervious were reclassified as impervious for the 30-m
classification.
This resulted in an overestimation of the impervious class but also
decreased the loss of information related to MRD due to resampling.
The NDVI-based change/no-change image was used to separate areas of change
and no-change from the 1990 Landsat image. Areas of no-change were used as training
for classification of the changed locations.
Data used for the 1990 classification
consisted of normalized data from all six Landsat reflective bands and derived indexes
50
(NDVI and TC).
classification.
The See5 data mining program (RuleQuest, 2008) was used for
See5 is a statistical data mining software package that performs CT
analysis with boosting. A total of 10 boosts were used. The unchanged 2005 classified
pixels were merged with the classified changed 1990 pixels to create a final 1990
classified map.
Accuracy for the 1990 classification was assessed using a stratified random
sampling design. A total of 302 sampling points were acquired. An additional 14 points
were manually identified because the stratified random design was unable to identify an
acceptable number of water points. These were then combined with the stratified random
points for a total of 316. These points were then visually compared to 1:40,000-scale
aerial photographs from the National Aerial Photography Program (NAPP) archives.
Producer’s and user’s accuracies for each class and the Kappa statistics were
calculated for each classification (Congalton and Green, 1999). Overall accuracy was
calculated using the methods outlined in Carrao et al., 2007. This method differs from
the traditional overall accuracy (Congalton and Green, 1999), in that it is calculated based
on the relative proportion of each class to the total number of classified pixels.
Our data set contained the full population of possible land cover change for our
study area.
This allowed descriptive statistics to be examined in order to evaluate
potential indicators of change in the study area. Four potential indicators were identified:
slope, aspect, distance-to-roads, and distance-to-streams. There were 16 theoretically
possible from-to change classes. Three of the 16 were identified as possibly being related
to MRD (Table 4.1). These were individually coded so as to separate the different types
51
of change for spatial pattern analysis. Two of the classes represent change from natural
habitats to human habitats through the conversion of vegetation to impervious surfaces.
Change class FI represented change from forests in 1990 to impervious surfaces 2005.
Change class GI represented change from grasslands/shrublands in 1990 to impervious
surfaces in 2005.
Change class FG represented changes in forests in 1990 to
grasslands/shrublands in 2005. A national forest boundary map created by the U.S.
Forest Service was used to exclude areas on forest service land where MRD was known
not to have occurred.
Table 4.1: Table of classification scheme for years 1990, 2005, and change image.
Classified as in 1990
Classified as in 2005
Change Class
Forests
Impervious Surfaces
FI
Grasslands/Shrublands
Impervious Surfaces
GI
Forests
Grasslands/Shrublands
FG
Slope and aspect were calculated from 1-m LiDAR elevation model resampled to
30 m using the nearest neighbor algorithm. Slope was calculated as degrees and aspect
was calculated as a categorical variable representing eight different directions (north,
northeast, east, southeast, south, southwest, west, and northwest) and flat. A distance-toroads layer was created by calculating distance from a road vector file of the area onto a
30-m grid. The road layer was created by the GIS Department of Gallatin County,
Montana, and contained all the roads in the study area. This layer is updated twice yearly
by Gallatin County and was current as of January 2008. Distance-to-streams represented
the topographically-driven flow path distance on a 30-m grid from streams identified by
using a single flow diration algorithm on a 10-m parsed elevation model created from the
52
original 1-m LiDAR elevation model. The 10-m parsed elevation model was created by
selecting every 10th point on the 1-m LiDAR model and rescaling to 10 m.
Descriptive statistics were computed relating each type of change to each
indicator variable being evaluated. The mean and standard deviation of each type of
change for each variable was calculated and compared to the mean and standard
deviation of land cover types and variables for the 1990 classification.
CT analysis was used to determine the relative relationship of the potential
indicators to the three different types of change. The three change classes were used as
the response variables. Slope, aspect, distance-to-streams, and distance-to-roads were
used as the explanatory variables. Standard cross validation methods were used to avoid
over-fitting the model (Venables and Ripley, 1997).
Results
Classification
The use of the generalized 2005 high-resolution Quickbird classification was
successful in the mapping of the 1990 Landsat image (Figure 4.2).
The 1990
classification had an overall accuracy of 86% with a Kappa statistic of 0.79 (Table 4.2).
All four classes had minimal error rates.
Grasslands and forests were most often
confused (Table 4.3). Water achieved no error of omission or commission. Impervious
surfaces had some confusion with grasslands.
53
Figure 4.2: Classified 1990 Landsat TM image based on 2005 Quickbird classification.
Impervious surfaces
Water
Forests
Overall accuracy = 86%
5
0
0
26
57
0
1
Forests
81
Water
Impervious
surfaces
Grasslands/Shrublands
Grasslands/
Shrublands
Table 4.2: Error matrix for 1990 Landsat TM classification. Columns represent
reference classes, while rows represent how the pixels were classified.
0
12
0
0
14
0
0
120
Kappa Statistic = 0.79
54
Table 4.3: User’s and producer’s accuracy for 1990 Landsat TM classification.
Class Name
Producer's Accuracy
User’s Accuracy
Grasslands/Shrublands
76%
83%
Impervious surfaces
Water
Forests
90%
100%
91%
100%
100%
82%
Change Detection
NDVI differencing successfully identified change related to MRD in our study
area (Figure 4.3). The NDVI differencing method captured patterns of development as
seen in the linear patterns and clustering of change pixels around the two major areas of
development in Big Sky, the Mountain Village and the Meadow Village. The NDVI
differencing method did not result in excessive amounts of false positive change
identification. The use of the forest boundary also aided in the minimization of false
positive change identification.
55
Figure 4.3: Classified NDVI difference image showing temporal from-to change classes.
Spatial Pattern Analysis
Analysis of from-to change indicated that the proportion of from-to changes were
disproportionate to the original land cover in 1990 (Table 4.4; Table 4.5). Forest change
accounted for 67% of the total change between 1990 and 2005. Forest, however, only
accounted for 51% of the land cover in 1990. Grassland changes accounted for 33% of
the total change and were proportional to its 1990 land cover with 29%.
56
Table 4.4: Percentage of each change class to the total amount of change between 19902005.
Change Class
Percent of Change
FI
48%
FG
19%
GI
33%
Table 4.5: Land cover percentages in 1990 Landsat classification.
Classified as in 1990
Percent of Land Cover in 1990
Grasslands
29%
Impervious Surface
19%
Forest
51%
Water
1%
The three change classes had differences among them for each of the indicator
variables. GI had the smallest slope mean followed by FG and FI changes (Table 4.6).
The three change classes differed in the mean response for distance-to-roads. GI had the
shortest distance followed by FI and FG. FI and FG changes had similar mean values for
distance to streams. .
Table 4.6: Mean value for indicator variables for from-to change change classes.
Change Class
Slope
Distance-to-Roads
Distance-to-Streams
FI
14
177.7
646
GI
9.4
90.3
397.3
FG
12.4
337.8
601
The proportion of 1990-2005 land cover changes to the overall proportion of land
cover classes in 1990 for each variable was also examined. Mean and standard deviation
for slope and grassland change did not differ from the values for grasslands in the 1990
classification (Table 4.7). FI and FG changes were combined in order to investigate
overall amount of forest loss. Forest changes occurred on a lower mean slopes than the
overall mean forest slopes in the 1990 classification.
57
Table 4.7: Table of mean and standard deviation values for the indicator variable slope
for forest and grasslands. Slope is represented in degrees.
Mean slope values in
Mean slope
Standard
Standard
1990 classification
value in change
deviation of
deviation of
classification
slope in 1990
slope in change
classification
classification
Forests
16.5
13
9.2
8.4
Grasslands
12.5
12.4
8.6
6.8
Mean changes in forests and grasslands were located closer to roads than the
mean response of forest and grasslands in the 1990 classification.
Both forest and
grassland changes averaged 300-m closer to roads than the mean forest and grassland
distance in 1990 classification. The standard deviations of forests and grasslands changes
were also smaller than the standard deviation of forest and grassland in the 1990
classification (Table 4.8).
Table 4.8: Table of mean and standard deviation values for the indicator variable
distance-to-roads for forest and grasslands. Distance-to-roads is represented in meters.
Mean distance-toMean distanceStandard
Standard
roads values in 1990 to-roads value
deviation of
deviation of
classification
in change
distance-todistance-toclassification
roads in 1990
roads in change
classification
classification
Forests
637.7
317
724.7
392.4
Grasslands
397
90.3
624.9
267.8
Changes in forest and grasslands differed in their mean response for distance-tostream (Table 4.9). Forested change was located farther way from streams than the mean
forest distance in the 1990 classification. Grassland change, however, was located closer
to streams than the mean grassland distance in the 1990 classification.
58
Table 4.9: Table of mean and standard deviation values for the indicator variable
distance-to-stream for forest and grasslands. Distance-to-stream is represented in meters.
Mean distance-toMean distanceStandard
Standard
stream values in
to-stream value
deviation of
deviation of
1990 classification
in change
distance-todistance-toclassification
stream in 1990
stream in
classification
change
classification
Forests
550.7
649.1
400.9
413.9
Grasslands
516.2
397.4
409.3
321.9
The relationship between forest and grassland conversion was also examined with
respect to aspect (Table 4.10). The overall proportion of forest change and the proportion
of forest in the 1990 classification to aspect were similar. The largest difference between
forest change and classified forest in 1990 was in the southeastern direction with 6% less
changed than represented in the 1990 classification. Grassland changes in aspect were
also similar to the proportion of grassland aspect in the 1990 classification. The largest
difference for grasslands was a 5% decrease in grasslands between 1990-2005 in the
northwest direction.
Table 4.10: Proportion of forest and grassland for nine different aspects to the overall
amount of forest and grassland in their classification.
Percent of
Percent of
Percent of
Percent of
forest change
forest class in
grassland
grassland class
between 19901990
change between
in 1990
2005
classification
1990-2005
classification
North
17%
15%
19%
15%
Northeast
20%
17%
19%
17%
East
14%
15%
12%
16%
Southeast
17%
11%
14%
15%
South
10%
10%
12%
13%
Southwest
6%
10%
7%
8%
West
5%
8%
4%
6%
Northwest
11%
14%
13%
8%
Flat
0.001%
0.001%
0.001%
0.001%
59
The CT based on the cross validation model was able to separate the three types of
change (Figure 4.4). The CT model accounted for 87% of the variance within this data set.
FI change was separated into four different groups. Three of the groups were located less
than 63.5 m from roads while one group was located greater than 63.5 m from roads. GI
change was grouped in to three groups, two being closer to roads than the last. All three
change classes occurred in areas greater than 63.5 m from roads. FG was captured in a
single group of change and was located greater than 63.5 m from roads.
The mapped classification tree displays the spatial patterns of from-to changes
(Figure 4.5). Blue shades represent GI changes, red shades represent FI changes, and green
represents FG change. Class A represents a dense cluster of GI change. This particular
group of change was caused by development of the area known as Meadow Village. FI
change due to the creation of ski slope is seen mostly in classes E and H. Class E, however
also capture the FI change of Mountain Village. The small amount of FG change can be
seen in an area with little development and can be attributed to forest management
practices such as logging.
60
Figure 4.4: Classification tree results. Branch length is proportional to amount of deviance
explained. End nodes represent the from-to change class. Nodes were coded
alphabetically in order to map their spatial distribution (Figure 4.5).
61
Figure 4.5: Classification tree based classified map of from-to changes. Each class is
coded alphabetically to match tree nodes from left to right (Figure 4.4). Blue shades
represent GI changes, red shades represent FI changes, and green represents FG change.
Discussion
Multi-resolution Image Classification
The successful use of the 2005 high-resolution Quickbird classification to map
historical land cover patterns in the 1990 Landsat image was largely due to the successful
identification of changed areas and the use of boosted CT. NDVI image differencing’s
ability to identify change is a testament to the method’s robustness in capturing vegetation
loss and tolerance of topographic variables. The speckled mountain slope indicates that
some false positive change could have been identified. These changes could be related to
62
difference in vegetation health, abundance, or structure between years, annual variation in
snow melt, and/or residual geometric errors.
NDVI differencing did not capture all differences between dates. The two images
used represent snapshots of the land’s surface at two moments in time. Forest areas that
could have been clear cut prior to 1990 and in the phase of regrowth in 1990 and in 2005
were not captured by the NDVI differencing method. This might be a deficiency of the
NDVI differencing method for vegetation monitoring purposes. Our study, however, was
not about forest management practices or vegetation dynamics. The appropriate method of
change identification is therefore related to the change one seeks to identify, and NDVI
differencing worked well for identifying MRD.
The success in the change identification resulted in the successful classification of
the 1990 Landsat image. The classification was largely based on the assumption that areas
with minimal change in NDVI were really areas of no change. A current classification
with a known accuracy and known land cover could then be used to map the historical
image.
Taking the classified changed areas and merging them with the current
classification reduced the compound error often associated with multiple classifications
(Howarth and Wickware, 1981). Compound error arises in independent classifications
because sources of error in the two classifications are largely independent. Compound
error can reduce the accuracy of the post classification comparison methods by multiplying
the error rates in each classification and greatly reducing the accuracy of the comparison.
Compound error is reduced in the approach used in this study by classifying a large portion
of the study area, which did not change between the two dates, only once.
63
Analysis of Temporal Change
Impervious surfaces encapsulated many different types of spectrally similar LULC
types in both classifications. Impervious surfaces represented the scree and talus present
on the mountain peaks and slopes, while at the same time representing human induced land
cover such as roads and residential and commercial development. The amalgamated land
cover that impervious surfaces represented was of limited utility for the individual
classifications, but did provide great insight into our temporal analysis because change
from vegetated to impervious surface was likely associated with increased MRD.
The 2005 classification had 1184.6 ha more area classified as impervious surfaces
than in the 1990 classification. This cannot be related to absolute increases in impervious
surfaces because the 30-m pixel size resulted in an abundance of pixels with mixed classes
and pixels were classified as impervious if 20% of the pixel contained impervious surfaces.
The 20% threshold was chosen in order to capture all impervious surfaces including small
narrow roads, which could have otherwise been excluded. The increase in impervious
surface was mirrored by a 944 ha decrease in forest and a 269.6 ha decrease in grasslands.
No catastrophic natural events occurred between the two dates that could have caused such
an increase. Forests and grasslands/shrublands conversion to impervious surfaces were
therefore likely a result of MRD induced land cover changes.
Results from the descriptive statistics indicated the conversion of forests and
grasslands to impervious surfaces was not proportional to their respective land cover in
1990. This is evident in the 67% of change being forest when forest only accounted for
51% of the land cover in 1990. This significant amount of forest loss can have negative
64
repercussions for ecosystems within the area. Most of the forest loss has a linear spatial
pattern. This is likely a result of the addition of ski runs to the landscape. These linear
patches differ from forest logging patches such as clear cuts in that the linear patches create
multiple smaller patches of forest. This results in more habitat fragmentation than would
have occurred with a clear cut. The effects of habitat fragmentation have been shown to
result in decreases in species composition, distribution, and abundance for both flora and
fauna (Odell et al., 2003; Crooks, 2002; Miller et al., 1998).
Previous studies have found topographic variables useful for identifying amenity
development (Cromartie and Wardell, 1999). Our topographic variables, aspect and slope,
did not provide much information on the proportion of change between 1990-2005 to land
cover in 1990. They did explain some of the variance in the CT model. Aspect was used
in the CT to separate two small groups of FI and FG. Slope was used to separate four
different groups of change in the CT.
The indicator variables, distance-to-streams and distance-to-roads, were effective at
explaining some of the variance within our change data. Forest and grassland changes, in
general, were found closer to roads than they were found in 1990. This, however, was at
least in part a result of roads being the actual change identified. MRD also generally
includes road development for access. Previous studies have found that the introduction of
roads to the landscape often facilitates development (Price, 2002). Accurate road data,
however, does not exist for 1990. We cannot, therefore, determine if this is true for our
case. We can use today’s roads, however, to identify possible area of future development.
65
The variable distance-to-stream provided good information on our types of change.
Previous studies have found that proximity of water is a good indicator of amenity
development (Parmenter et al., 2003; Cromartie and Wardell, 1999). Our data showed that
development of grasslands was disproportionately closer to streams than grasslands in
general. This could likely be a result of amenity development at or near the water edges
(Williams and McMillan, 1983; Williams and Jobes, 1990). The proximity of Meadow
Village, a major focus of development, to the West Fork of the Gallatin River, certainly
accounts for a large portion of this change. Forest change, however, was found farther
away than the mean forest land in 1990. These relationships could also be a result of
grasslands naturally occurring closer to water on average than forested land, which often
occurs on upland mountain slopes.
Conclusion
High-resolution imagery was successfully merged with moderate-resolution
imagery to map changes in land cover patterns in Big Sky, Montana. Previous research in
this area has been lacking.
Our research indicated that the generalization of a high-
resolution classification can be used as training data for a historical image. This approach
shows promise as a method of temporally monitoring LULC changes.
The use of NDVI image differencing and boosted CT resulted in the successful
identification of land use change due to MRD. The NDVI differencing method allowed
only areas of change to be identified and reclassified. This decreased the compound error
often associated with post classification comparisons.
Boosted CT handled both the
66
Landsat data and derived vegetation indices well. This is largely a result of the robustness
of the classifier. Future research should include mapping more than two dates. This can
allow for more information on the rate and nature of changes between years.
Statistical pattern analysis demonstrated that our indicators could be used to explain
the change within our study area. We found forest changes to have a disproportionate
amount of change when compared to the overall amount of forested area in 1990. We also
found that forest changes were located farther away from streams and tended to occur on
lower slopes. Grassland change tended to occurr closer to steams. Overall changes in
grasslands was proportional to the 1990 amount of land cover. These indicator variables
explained 87% of the variance for the change classes and might be related to amenity
development.
67
CHAPTER 5
CONCLUSION
Mountain Resort Development (MRD) is rapidly changing the landscape of the
intermountain west. The emerging pattern is a shift from a primary extractive economy
to a tertiary economy.
The result of MRD is changes in LULC, which can affect
ecosystems within the developed area, adjoining undeveloped areas, and downstream and
riparian systems. Big Sky, Montana, provided a valuable opportunity to assess current
and historical LULC patterns associated with MRD. Current LULC patterns in Big Sky
were identified by creating a detailed classification using high-resolution imagery
(multispectral and LiDAR) with object-oriented analysis. Historical land cover patterns
were analyzed using change detection methods and descriptive statistical methods.
Results of this study indicated object-oriented analysis and classification was an
effective means of creating detailed LULC maps with high-resolution imagery. Objectoriented classification has the ability to overcome the spectral limitations of highresolution sensors through the addition of contextual metrics to the classification process.
This creates additional dimensions for separating spectrally similar classes.
The
classified LULC maps in this study based on Quickbird imagery had an overall accuracy
of 90% for both levels of the classification hierarchy. This exceeded USGS standards for
remotely sensed classifications.
Image fusion in the form of combined multispectral information and elevation
information derived from LiDAR first and last returns created more realistic objects than
68
multispectral objects alone.
Object-oriented classification of a fused Quickbird and
LiDAR image resulted in reduced error rates for individual classes compared to the
Quickbird classification.
The two classifications were not, however, statistically
different. The overall accuracy for the fused image was 91% for both levels of the
classification hierarchy.
The fused classification resulted in a visually appealing
classification with objects that better resembled their real world land cover classes. This
was a result of a finer segmentation produced by the addition of the LiDAR data to the
classification.
Future research opportunities for object-oriented classification and analysis are
numerous. The identification of the appropriate spatial resolution for use with objectoriented classification is essential. This might prevent the creation of mixed land cover
objects caused by the pixilated effect seen with the Quickbird imagery. Research into the
full utilization of contextual metrics offered by the object-oriented software is also
recommended. Object-oriented software also needs to be programmed to enable efficient
handling of large images so as to make full use of the today’s high resolution data.
Temporal analysis of land cover patterns was accomplished by successfully using a
generalized version of the high-resolution land cover map as baseline data combined with
NDVI image differencing based on Landsat imagery. This essentially resulted in the
calibration of the historical Landsat image based on the Quickbird and LiDAR objectoriented classification. Previous research on multitemporal mapping of multiresolution
images has been lacking.
Our research indicated that the generalization of a high-
69
resolution classification can be used as training data for a historical image. This shows
promise as a method of temporally monitoring LULC changes.
NDVI image differencing and boosted classification trees resulted in the successful
identification of temporal changes in land cover due to MRD. The NDVI differencing
method successfully identified the land cover change due to vegetation loss and was not
affected by the topography of the study area. This allowed for the identification and
classification of change areas only. The classified change areas were then merged with the
unchanged current land cover classification. The merging process effectively decreased the
compound error associated with multiple classifications.
Numerous future research possibilities exist for multitemporal mapping of
multiresolution images. Mapping more than two dates is recommended. This can allow
for more information on the rate and nature of changes between years to be gleaned. The
land cover classes used for temporal analysis in our study were broad and general. An
attempt should be made to identify how fine a classification can be generated from the high
resolution classification.
Statistical pattern analysis demonstrated that distance-to-streams, distance-to-roads,
slope, and aspect were all correlated to change within our study area. Forest changes were
found to be disproportionate to their landscape amounts in 1990. Forest changes were also
found to be located farther away from streams and on lower slopes than their 1990
proportions. Grassland change tended to occurred closer to steams. Overall changes in
grasslands, however, were proportional their 1990 land cover. These variables explained
87% of the variance for the change classes and might be related to amenity development.
70
The statistical pattern analysis also indicated an increase in impervious surfaces
between the years 1990-2005. This increase in impervious surface resulted in a decrease in
both forests and grassland areas.
Loss of forest and grassland area increases habitat
fragmentation and can have negative consequences for ecosystems within the area.
71
LITERATURE CITED
Allen, T. R., and Kupfer, J. A., 2000. Application of spherical statistics to change vector
analysis of Landsat data southern Appalachian spruce-fir forests. Remote Sensing
of Environment 74:482-493.
Baker, C., Lawrence, R., Montagne, C., and Patten, D., 2007. Change detection of
wetland and riparian ecosystems using change vector analysis. Wetlands 27:610619.
Baker, C., Lawrence, R., Montagne, C., and Patten, D., 2006. Mapping wetlands and
riparian areas using Landsat ETM+ imagery and decision tree-based models.
Wetlands 26:465-474.
Baron, J. S., Theobald, D. M., and Fagre, D. B., 2000. Management of land use conflicts
in the United States Rocky Mountains. Mountain Research and Development
20:24-27.
Benz, U. C., Hofmann, P., Willhauck, G., Lingenfelder, I., and Heynen, M., 2004. Multiresolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready
information. ISPRS Journal of Photogrammetry and Remote Sensing 58:239-256.
Bricklemyer, R.S., Lawrence, R.L., Miller, P.R., and Battogtokh, N., 2007. Monitoring
and verifying agricultural practices related to soil carbon sequestration.
Agriculture, Ecosystems and Environment 118:201-210.
Brown, K., Turner, R. K., Hameed, H., and Bateman, I., 1997. Environmental carrying
capacity and tourism development in the Maldives and Nepal. Environmental
Conservation 24:316-325.
Budreski, K. A., Wynne, R. H., Browder, J. O., and Campbell, J. B., 2007. Comparison
of segment and pixel-based non-parametric land cover classification in the
Brazilian Amazon using multitemporal Landsat TM/ETM+ imagery.
Photogrammetric Engineering and Remote Sensing 73:813-827.
Carleer, A. P., and Wolff, E., 2006. Urban land cover multi-level region-based
classification of VHR data by selecting relevant features. International Journal of
Remote Sensing 27:1035-1051.
Carrao, H., Caetano, M., and Coelho, P. S., 2007. Sample design and analysis for
thematic map accuracy assessment: An approach based on domain estimation for
the validation of land cover products. Proceedings of 32nd International
Symposium on Remote Sensing of Environment, Costa Rica June 25-29, 2007.
72
Chubey, M. S., Franklin, S. E., and Wulder, M. A., 2006. Object-based analysis of
Ikonos-2 imagery for extraction of forest inventory parameters. Photogrammetric
Engineering & Remote Sensing 72:383-394.
Collins, J. B., and Woodcock, C. E., 1996. An assessment of several linear change
detection techniques for mapping forest morality using multitemporal Landsat
TM data. Remote Sensing of Enviroment 1: 66-77.
Congalton R. C., and Green, K., 1999. Assessing the accuracy of remotely sensed data:
Principles and Practices. Lewis Publishers, New York.
Coppin, P., Jonckheere, I., Nackaerts, K., and Muys, B., 2004. Digital change detection
methods in ecosystem monitoring: A review. International Journal of Remote
Sensing 19:411-426.
Crist, E. P., and Cicone, R. C., 1984. A physically-based transformation of Thematic
Mapper data the TM tasseled cap. IEEE Transactions of Geosciences and Remote
Sensing 25:1565-1596.
Cromartie J. B., and Wardwell J. M., 1999. Migrants settling far and wide in the rural
West. Rural Development Perspectives 14:2-8.
Crooks, K. R., 2002. Relative Sensitivities of Mammalian Carnivores to Habitat
Fragmentation. Conservation Biology 16:488-502.
De’ath, G., and Fabricius, K. E., 2002. Classification and Regression Trees: A powerful
yet simple technique for ecological data analysis. Ecology 81:3178-3192.
Definiens Professional 5: LDH, 2007. User’s Guide.
Definiens Professional 5: LDH, 2007. Reference Manual.
Friedl, M. A., and Brodley, C. E., 1997. Decision tree classification of land cover from
remotely sensed data. Remote Sensing of Environment 61:399-409.
Gardener, K. K., and Vogel, R. M., 2005. Predicting groundwater nitrate concentrations
from land use. Groundwater 43:343-352.
Green, K., and Lopez, C., 2007. Using Object-Oriented Classification of ADS40 to map
benthic habitats of the state of Texas. Photogrammetric Engineering and Remote
Sensing 73:861-865.
73
Groffman, P. M., Bain, D. J., Band, L. E., Belt, K. T., Brush, G. S., Grove, J. M., Pouyat,
R. V., Yesilonis, I. C., Zipperer, W. C., 2003. Down by the riverside: Urban
riparian ecology. Frontier in Ecology and the Envirionment. 6:315-321.
Guerra, F., Puig, H., and Chaume, R., 1998. The forest-savanna dynamics from multidate Landsat TM in Sierra Parima, Venezuela. International Journal of Remote
Sensing 19:2061- 2075.
Haala, N., and Brenner, C., 1999. Extraction of buildings and trees in urban
environments. ISPRS Journal of Photogrammetry & Remote Sensing 54:130-137.
Hansen, A. J., Rasker, R., Maxwell, B., Rotella, J. J., Johnson, J. D., Wright Parmenter,
A., Langer, U., Cohen, W. B., Lawrence, R. L., and Kraska, M. P.V., 2002.
Ecological causes and consequences of demographic change in the New West.
Bioscience 52:151-162.
Herold, M., Gardner, M. E., and Roberts, D. A., 2003. Spectral resolution requirements
for mapping urban areas. IEEE Transactions on Geoscience and Remote Sensing
41:1907-1919.
Herold, M., Roberts, D. A., Gardner, M. E., and Dennison, P. E., 2004. Spectrometry for
urban area remote sensing–development and analysis of a spectral library from
350 to 2400 nm. Remote Sensing of Environment 91:304-319.
Herold, M., and Scepan, J., Muller, A., and Gunther, S., 2002. Object-oriented mapping
and analysis of urban land use/cover using IKONOS data. Proceedings of 22nd
EARSEL Symposium for European-wide integration, Prague.
Herold, M., Gardner, M. E., and Roberts, D. A., 2003. Spectral resolution requirements
for mapping urban areas. IEEE Transactions on Geoscience and Remote Sensing
41:1907-1919.
Hill, R. A., and Thomson, A. G., 2005. Mapping woodland species composition and
structure using airborne spectral and LiDAR data. International Journal of
Remote Sensing 26:3763-3779.
Hyde, P., Dubaya, R., Walker, W., Blair, J. B., Holten, M., and Hunsaker, C., 2006.
Mapping forest structure for wildlife habitat analysis using multi-sensor (LiDAR,
SAR/InSAR, ETM+, Quickbird) Synergy. Remote Sensing of Environment
102:63-73.
Howarth, P. J., and Wickware, G. M., 1981. Procedures for change detection using
Landsat. International Journal of Remote Sensing 12:1471-1491.
74
Johnson, R. D., and Kasischke, E. S., 1998. Change vector analysis: A technique for the
multi-spectral monitoring of land cover and condition. International Journal of
Remote Sensing 19:411-426.
Kuehn, S., Benz, U., and Hurley, J., 2002. Efficient flood monitoring based on
RADARSAT-1 images data and information fusion with object-oriented
technology. IEEE Transactions on Geoscience and Remote Sensing Symposium.
Laliberte, A. S., Rango, A., Havstad, K. M., Paris, J. F., Beck, R. F., McNeely, R., and
Gonzalez, A. L., 2004. Object-oriented image analysis for mapping shrub
encroachment from 1937 to 2003 in Southern New Mexico. Remote Sensing of
Environment 93:198-210.
Lawrence, R., Bunn, A., Powell, S., Zambon, M., 2004. Classification of remotely sensed
imagery using stochastic gradient boosting as a refinement of classification tree
analysis. Remote Sensing of Environment 3:331-336.
Li, X., and Yeh, A. G. O., 1998. Principal component analysis of stacked multi-temporal
images for the monitoring of rapid urban expansion in the Pearl River Delta.
International Journal of Remote Sensing 19:1501-1518.
Liu, X., and Lathrop, R. G., 2002. Urban change detection based on an artificial neural
network. International Journal of Remote Sensing 23:2513-2518.
Lu, D., Mausel, P., Brondizio, E., and Moran, E., 2003. Change detection techniques.
International Journal of Remote Sensing 25:2365-2407.
Lynn, N. A., and Brown, R. D., 2003. Effects of recreational use impacts on hiking
experiences in natural areas. Landscape and Urban Planning 64:77-87.
Lyon, J. G., Yuan, D., Lunetta, R. S., Elvidge, C. D., 1998. A change detection
experiment using vegetation indices. Photogrammetric Engineering and Remote
Sensing 64:143-150.
Marston, Richard A., and Anderson, J. E., 1991. Watersheds and vegetation of the
Greater Yellowstone Ecosystem. Conservation Biology 5:338-346.
Mas, J. F., 1999. Monitoring land-cover changes: A comparison of change detection
techniques. International Journal of Remote Sensing 20:139-152.
McCombs, J. W., Roberts, S. D., and Evans, D. L., 2003. Influence of fusing LiDAR and
multi-spectral imagery on remotely sensed estimates of stand density and mean
tree height in a managed loblolly pine plantation. Forest Science 49:457-466.
75
Miller, S. G., Knight, R. L., and Miller, C. K., 1998. Influence of recreational trails on
breeding bird communities. Ecological Applications 8:162-169.
Mitri, G. H., and Gitas, I. Z., 2004. A performance evaluation of a burned area objectbased classification model when applied to topographically and nontopographically corrected TM imagery. International Journal of Remote Sensing
25:2863-2870.
Moeller, M., and Blaschke, T., 2006. A new index for the differentiation of vegetation
fractions in urban neighborhoods based on satellite imagery. ASPRS Annual
Conference, Reno Nevada 2006.
Odell, E. A., Theobald, D. M., and Knight, R. L., 2003. Incorporating ecology into land
use planning: The songbirds’ case for clustered development. Journal of
American Planning and Association 69:72-82.
Parmenter Wright, A., Hansen, A., Kennedy, R. E., Cohen, W., Langner, U., Lawrence,
R., Maxwell, B., Gallant, A., and Aspinall, R., 2003. Land use and land cover
change in the Greater Yellowstone Ecosystem:1975-1995. Ecological
Applications 13:687-703.
Pohl, C., and Van Genderen, J. L., 1998. Multi-sensor image fusion in remote sensing:
Concepts, methods and applications. International Journal of Remote Sensing
19:823-854.
Price, M. F., 1992. Patterns of the development of tourism in mountain environments.
GeoJournal 27:87-96.
Reed, R. A., Johnson-Barnard, J., and Baker, W., 1996. Contributions of roads to forest
fragmentation in the Rocky Mountains. Conservation Biology 10:1098-1106.
Ridd, M. K., and Liu, J., 1998. A comparison of four algorithms for change detection in
an urban environment. Remote Sensing of Environment 63:95-100.
Riebsame, W. E., Gosnell, H., and Theobald, D. M., 1996. Land Use and Landscape
change in the Colorado mountains 1: Theory, scale, pattern. Mountain Research
and Development 16:395-405.
Ries, J., 1996. Landscape damage by skiing at the Schauinsland in the Black Forest,
Germany. Mountain Research and Development 16:27-40.
Rouse, J. W., Haas, R. H., Schell, J. A., Deering, D. W., and Harlan, J. C., 1973.
Monitoring the vernal advancement of retrogradation (green wave effect) of
natural vegetation. NASA/GSFC Type III Final Report, Greenbelt, MD.
76
RuleQuest Research, 2008. Data Mining Tools See5 and C5.0 2.05.
Sauvajot, R. M., Buechner, M., and Kamradt, C. M., 1998. Patterns of human disturbance
and response by small mammals and birds in chaparral near urban development.
Urban Ecosystems 2:279-297.
Sawaya, K. E., Olmanson, L, G., Heinert, N, J., Brezonik, P. L., and Baur, M. E., 2003.
Extending satellite remote sensing to local scales: Land and water resource
monitoring using high-resolution imagery. Remote Sensing of Environment
88:144-156.
Shanley, J. S., and Wemple, B. C., 2002. Water quality and quantity in the mountain
environment. Vermont Law Review 26 (3):717-751.
Shiba, M., and Itaya, A., 2006. Using eCognition for improved forest management and
monitoring systems in precision forestry. Proceeding of the International Recision
Forestry Symposium, Stellenbosch University.
Venables, W. N., and Ripley, B. D., 1997. Modern Applied Statistic with S-plus Second
Edition, Springer-Verlas, New York, NY, 5489.
Walker, J. S., and Briggs, J. M., 2007. An object-oriented approach to urban forest
mapping in Phoenix. Photogrammetric Engineering and Remote Sensing 73:577583.
Wang, L., Sousa, W. P., and Gong, P., 2004. Integration of object-based and pixel-based
classification for mapping mangroves with IKONOS imagery. International
Journal of Remote Sensing 20:5655-5668.
Weaver, J. L., Paquet, P. C., and Ruggiero, L. F., 1996. Resilience and conservation of
large carnivores in the Rocky Mountains. Conservation Biology 10:964-976.
Wemple, B. C., Jones, J. A., and Grant, G. E., 1996. Channel network extension by
logging roads in two basins, Western Cascades, Oregon. Water Resources
Bulletin 32:1195-1207.
Wemple, B. C., Swanson, F. J., and Jones, J. A., 2001. Forested roads and geomorphic
process interactions, Cascade Range, Oregon. Earth Surface Processes and
Landforms 26:191-204.
Williams, A. S., and McMillan, D. B., 1983. Location specific capital and destination
selection among migrants to non-metro areas. Rural Sociology 48:457-497.
77
Williams, A. S., and Jobes, P. C., 1990. Economic and quality of life considerations in
urban-rural migration. Journal of Rural Studies 6:187-194.
Yuan, F., and Bauer, M., 2006. Mapping impervious surface area using high resolution
imagery A comparison of object-based and per pixel classification. ASPRS
Annual Conference May 20